ArticlePDF Available

Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension

Authors:

Abstract

Self-paced reading and speeded acceptability judgement experiments on Japanese sentence comprehension were conducted to elucidate the processing mechanisms of ambiguity resolution in structural reanalysis. The results suggest that the parser attempts to maintain an initial clause-mate relationship of noun phrases in the process of reanalysis, which is consistent with the prediction of the Minimal Revision Principle (Frazier and Clifton, 1998), but not with that of the Minimum Maximal-Projection Principle (Oishi, 2007).
Structural ambiguity resolution in the process
of reanalysis: Evidence from Japanese sentence
comprehension
著者 KUSABA SUZUE, YANO MASATAKA
journal or
publication title Tohoku psychologica folia
volume 76
page range 20-37
year 2018-03-30
URL http://hdl.handle.net/10097/00122647
Tohoku Psychologica Folia
2017, 76, 20−37
Self-paced reading and speeded acceptability judgement experiments on Japanese sentence
comprehension were conducted to elucidate the processing mechanisms of ambiguity resolution
in structural reanalysis. The results suggest that the parser attempts to maintain an initial
clause-mate relationship of noun phrases in the process of reanalysis, which is consistent with
the prediction of the Minimal Revision Principle (Frazier & Clifton, 1998), but not with that of
the Minimum Maximal-Projection Principle (Oishi, 2007).
Key words: sentence comprehension, ambiguity resolution, garden-path sentences,
reanalysis, Japanese
Introduction
Psycholinguistic research on sentence comprehension has been concerned with how the
language processing mechanism works in the brain. Traditional studies, focusing primarily
on syntactic processing, have found that people parse continuous input incrementally, even in
verb-final languages. For example, readers routinely experience processing difficulties when
a temporarily ambiguous sentence is disambiguated in favour of a less preferred structure,
such as in the case of The horse raced past the barn fell’ (Bever, 1970). The processing difculty
increases in such a sentence because readers process ‘raced’ as a main verb in the past tense
without waiting for the later information that signals that ‘raced past the barn’ is a reduced
relative clause (meaning ‘the horse that was raced past the barn’). Therefore, readers struggle
to parse the structure correctly with ‘fell’ occurring at the end. This extra processing cost
is referred to as a garden-path (GP) effect. The GP effect indicates that people incorporate
encountered words into the syntactic representation by resolving a structural ambiguity
without a signicant delay. The GP effect provides useful information to examine a parsing
mechanism in the initial analysis (e.g., Frazier, 1979; Frazier & Clifton, 1986; Frazier & Fodor,
1978).
Relatively less known is the mechanism of ambiguity resolution in reanalysis. Upon
encountering an error signal (a word that triggers reanalysis), there are sometimes multiple
structural possibilities to be adopted, especially in head-nal languages such as Japanese. For
example, native speakers of Japanese prefer to initially analyse the sentence in (1a) below
as a simple sentence, although more syntactically complex structures are available (Inoue,
1. School of Letters, Kyushu University
2. Graduate School of Arts and Letters, Tohoku University
3. Japan Society for the Promotion of Science (JSPS)
Structural ambiguity resolution in the process of reanalysis:
Evidence from Japanese sentence comprehension
SUZUE KUSABA
(草場鈴恵)
1 and MASATAKA YANO
(矢野雅貴)
2, 3
21
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
2006; Mazuka & Itoh, 1995). Since the relative clause in Japanese is pre-nominal and does not
have a clause boundary marker, such as ‘who’ in English (e.g., ‘The girl who observed a child’),
Japanese speakers experience a GP effect upon encountering the head noun of the relative
clause, such as ‘girl-dat’ in (1b). Consequently, they are required to revise a syntactic structure
to incorporate it into an existing structure in a legitimate way. Importantly, there are (at least)
two possible structures that the parser can build, as indicated in (2). In (2a), the accusative
NP ‘child-acc’ is analysed as a constituent in the relative clause headed by the error signal
girl-dat’, whereas it exists in the matrix clause in (2b). Mazuka and Itoh (1995) have argued
that Japanese speakers prefer (2a) to (2b) in resolving structural ambiguity on the basis of
the intuitive judgement that (3a), which is compatible with the structure in (2a), is easier to
comprehend than (3b), in which an inanimate NP ‘taxi-dat’ allows only for the structure in
(2b) (see also Hirose & Inoue, 1998). Mazuka and Itoh (1995) suggest that as the number of
constituents that are reanalysed as belonging to a different clause than the relative clause
increases, such a structure is not preferred.
(1) a. Yoko-ga kodomo-o koosaten-de mikaketa
Yoko-nom child-acc intersection-loc saw
‘Yoko saw a child at the intersection’
b. Yoko-ga kodomo-o koosaten-de mikaketa onnanoko-ni ...
Yoko-nom child-acc intersection-loc saw girl-dat
(2) a. [Yoko-ga [[S-gapi kodomo-o koosaten-de mikaketa] onnanokoi-ni] (V) ]
Yoko-nom child-acc intersection-loc saw girl-dat
‘Yoko (V) the girl who saw a child at the intersection.’
b.
[Yokoi-ga [kodomo-o [S-gapi/j/k O-gapi/j/k koosaten-de mikaketa] onnanokoj-ni] (V) ]
Yoko-nom child-acc intersection-loc saw girl-dat
‘Yokoi (V) the child on/to the girl {who shei/someone saw} at the intersection.’
‘Yokoi (V) the child on/to the girl {who saw heri/someone} at the intersection.’
(3) a.
[Yoko-ga [[S-gapi kodomo-o koosaten-de mikaketa] onnnanokonii-ni] koe-o kaketa.]
Yoko-nom child-acc intersection-loc saw girl-dat called.
‘Yoko called the girl who saw a child at the intersection.’
b.
[Yokoi-ga [kodomo-o [S-gapi O-gapj koosaten-de mikaketa] takushiij-ni] noseta.]
Yoko-nom child-acc intersection-loc saw taxi-dat put.in
‘Yokoi put the child in the taxi that shei saw at the intersection.’
Minimal Revision
Mazuka and Itoh’s (1995) observation can be accounted for by the Minimal Revisions
Principle (MRP), as stated in (4). This principle requires the parser to maintain an initial
analysis to the extent possible. Accordingly, this principle successfully predicts that the parser
22 Kusaba, S. and Yano, M.
should select the structure in (2a), in which only the nominative subject belongs to a different
clause than other constituents after revising an initial structure (see also Sturt & Crocker, 1996
for a different proposal).
(4) Minimal Revisions (MR): Do not make an unnecessary revision. When revision is
necessary, make the minimal revision consistent with the error signal and maintain as
much of the already assigned structure and interpretation as possible.
Minimum Maximal-Projection Principle
Oishi (2007) identified several problems with the MRP. First, although this principle
stipulates what structure the parser should select, it does not stipulate how the parser does
so. Under this principle, the parser might compute all possible structures and choose the best
structure in accordance with the principle, but this is unlikely in terms of the limitation of
cognitive resources (Oishi, 2007: 33). Second, it is not clear how to analyse a syntactic structure
when the error signal is encountered before completion of an initial analysis (i.e., simple clause)
(Oishi, 2007: 34–35). In (5), a simple clause analysis is semantically incongruent after reading
a verb, signalling that this analysis is unlikely. Thus, the parser has to posit a gap inside the
relative clause with or without ‘secretary-dat’, as shown in (6). According to Oishi (2007), the
MP principle cannot make an explicit prediction with respect to which structure the parser
builds in the case of (5).
(5)
?
daijin-ga hisho-ni atsumatta
minister-nom secretary-dat gathered
Lit. ‘The minister gathered at the secretary.’
(6) a. daijin-ga [S-gapi hisho-ni atsumatta] NPi ...
minister-nom secretary-dat gathered
b. daijin-ga hisho-ni [S-gapi atsumatta] NPi ...
minister-nom secretary-dat gathered
Oishi (2007) proposed an alternative principle that he referred to as the Minimum
Maximal-Projection Principle (MMPP), as dened in (7), to solve the problems of the MRP.
In the case of (2), the MMPP expects that the parser will select the structure in (2a) because
(2a) has fewer gaps than (2b) in the relative clause. Furthermore, since it does not require
the parser to maintain an initial analysis, this principle may potentially explain structural
ambiguity resolution preferences, such as in the case of (5), wherein the verb induces a
reanalysis process even before completion of a simple clause analysis. The MMPP predicts that
the parser will prefer (6b) to (6a) because the number of nodes in the relative clause is fewer in
(6b) than (6a) while the number of gaps is the same between (6a) than (6b).
2
23
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
(7) Minimum Maximal-Projection Principle (MMPP):
When an error signal is encountered, construct the minimum maximal-projection
that has the fewest number of nodes among those that can dominate the error
signal and can be attached into the existing structure in a legitimate way such that
grammatical rules are not violated. Fill each argument position in the minimum
maximal-projection with an overt element if possible. (Oishi, Yasunaga, & Sakamoto,
2007)
Oishi (2007) conducted a self-paced reading experiment to test the MMPP using four types
of sentences in (8), in which animacy of the dative object (animate vs. inanimate) and type of
matrix verb (ditransitive vs. monotransitive) were crossed. (The error signals were underlined.)
(8) a. Animate/ditransitive:
akumyoodakai daijin-ga jitoo-no hisho-ni atsumatta uragane-o azuketa.
notorious minister-nom
own.party-gen
secretary-dat
gathered bribe-acc entrusted
‘The notorious minister entrusted the bribe that was disbursed to the secretary of his own party.’
b. Animate/monotransitive:
akumyoodakai daijin-ga jitoo-no hisho-ni atsumatta uragane-o nusunda.
notorious minister-nom
own.party-gen
secretary-dat
gathered bribe-acc stole.
The notorious minister stole the bribe that was disbursed to the secretary of his own party.’
c. Inanimate/ditransitive:
akumyoodakai daijin-ga jitoo-no honbu-ni atsumatta uragane-o azuketa.
notorious
minister-nom own.party-gen headquarters-dat
gathered bribe-acc entrusted
‘The notorious minister entrusted the bribe that was disbursed to the headquarters of his own party.’
d. Inanimate/monotransitive:
akumyoodakai daijin-ga jitoo-no hisho-ni atsumatta uragane-o nusunda.
notorious
minister-nom
own.party-gen
headquarters-dat
gathered bribe-acc stole.
‘The notorious minister stole the bribe that was disbursed to the headquarters of his own party.’
If the parser follows the MRP, it should build a monotransitive structure including an
accusative object with a relative clause (i.e., [the bribe that was disbursed to the secretary/
headquarters of his own party]) after encountering an error signal because this structure
allows it to maintain an initial analysis ofsecretary/headquarters-dat’ as an NP co-occurring
within the same clause as the verb ‘gathered’. Thus, this principle predicts a longer reading time
for the ditransitive condition than the monotransitive condition, reecting a second reanalysis
2. One may think that MMPP should predict a preference for (6a) because the dative NP position is
occupied by the overt phrase ‘secretary-dat’ in (6a) but not in (6b). Thus, the second condition of MMPP
(‘Fill each argument position in the minimum maximal-projection with an overt element if possible’) forces
the parser to select a structure in (6a). However, Oishi (2007) assumes that the dative NP phrase is an
adjunct and that there is no gap for this adjunct when ‘secretary-dat’ is interpreted outside of the relative
clause.
24 Kusaba, S. and Yano, M.
cost.
3
On the other hand, the MMPP offers an opposite prediction regarding a processing
asymmetry between the monotransitive and ditransitive verbs. This principle predicts that the
monotransitive verb should incur a larger processing cost than the ditransitive verb because
the ditransitive structure with fewer nodes inside the relative clause should be selected when
encountering an error signal. The result of the self-paced reading experiment showed a longer
reading time for the monotransitive verb than the ditransitive verb, irrespective of the input
timing of the error signal (i.e., the manipulation of animacy), thus supporting the MMPP.
4
To summarise, two principles have been proposed to account for ambiguity resolution
preference in reanalysis, namely, the MRP and the MMPP. The MMPP has been borne out
by Oishi’s (2007) observation that the parser has a strategy for analysing a dative NP as a
constituent outside the relative clause rather than inside the relative clause.
The present study aims to test the MMPP using transitive sentences wherein the thematic
relationship between co-arguments differs from that of intransitive sentences in the initial
analysis. In Oishi’s (2007) experiment, the parser has two options as to the position of a
left clause boundary of the relative clause, as indicated by ‘[’ in (9) below. As discussed
above, Oishi (2007) showed that the parser selects the second boundary immediately before
the verb upon encountering an error signal. Although Oishi (2007) explained this result by
proposing the MMPP, it might be accounted for by a preference for not abandoning a close
relationship between the nominative subject ‘minister-nom’ and the dative object ‘headquarters-
dat’, which are initially analysed as the THEME and GOAL of ‘gathered vi’, respectively.
5
In other words, the parser initially interprets this sentence as ‘The minister(s) move to the
headquarters’ and later reanalyses as ‘The minister(s) move to the headquarters with the bribe’,
expecting a verb such as ‘hakonda’ (brought) to avoid abandoning a relationship between
the THEME ‘minister’ and the GOAL ‘headquarters’. If this possibility is on the right track,
we expect a different preference in the transitive sentence, such as that exemplied in (10).
The parser should initially analyse this sentence as ‘The minister causes the bribe to move to the
headquarters’ before encountering an error signal. In other words, ‘headquarters-dat’ is a GOAL
and ‘bribe-acc’ is a THEME of the rst verb. If the parser prefers to maintain the clause-mate
relationship between a THEME phrase and a GOAL phrase even after reanalysis, it should
insert a left clause boundary at the second position rather than the third position to include
the GOAL phrase ‘headquarters-dat’ inside the relative clause. This prediction is the same as
3. At the end of a sentence, there are two possible ways to revise a monotransitive structure into a
ditransitive one. The parser can posit a pro in the GOAL position of entrusted’. Alternatively, the parser
can reanalyse the dative NP ‘secretary/headquarters-dat’ as a GOAL of entrusted’. In the latter case,
however, the parser has to abandon a structural relationship between the dative NP and the rst verb,
and thus this revision is less likely under the MRP.
4. Oishi (2007: 43) reported no significant effect of ANIMACY at the fifth region (i.e., the verb of the
relative clause) or the sixth phrase (i.e., the head noun of the relative clause) (see Oishi, 2007: 44 and
Experiment 2 for discussion).
5. Here, we assume that the subject of atsumatta’ (gathered vi) is a THEME argument because the subject
of this verb can be inanimate, as shown in (i).
(i) uragane-ga honbu-ni atsumatta.
bribe-nom headquarters-dat gathered.
‘The bribe was disbursed to the headquarters.’
25
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
that predicted by the MRP. In contrast, the MMPP predicts that the parser should adopt
the third clause boundary by taking ‘headquarters-dat’ out of the clause with ‘bribe-acc’ to
make the maximal projection as small as possible. We conducted two behavioural experiments
using similar sentences as in (10), where our hypothesis/MRP and the MMPP offer different
predictions.
(9) Oishi (2007):
daijin-ga [honbu-ni [atsumatta uragane-o ...
minister-nom headquarters-dat gathered vi bribe-acc
(10) [daijin-ga [honbu-ni [uragane-o atsumeta hisho-o ...
minister-nom headquarters-dat bribe-acc gathered
vt secretary-acc
Experiment 1
The purpose of this experiment is to examine the parser’s preference in resolving
structural ambiguity during the process of reanalysis using a self-paced reading task.
Specically, we examined whether the MRP and MMPP correctly predict a GP effect induced
by the parser’s prior disambiguation process.
Stimuli
We created four types of experimental sentences, such as those listed in (11), in which the
presence of GP (GP/non-GP) and VERB type (monotransitive/ditransitive) were manipulated.
The GP conditions comprised six phrases, whereas the non-GP conditions consisted of
four phrases. In the fifth region of the GP conditions, the parser was required to revise a
structure to incorporate a (head) noun into an existing structure. The region of interest in
this experiment centred on the matrix verbs at the end of a sentence (underlined), which were
matched between monotransitive and ditransitive verbs on the number of characters (t (18)
= 0.001, p > 0.10), the number of morae (t (18) = 0.22, p > 0.10), and auditory/visual word
familiarity taken from the Lexical Properties of Japanese database (Amano & Kondo, 1999)
(t (18) = 0.24, p > 0.10).
(11) a. GP/monotransitive:
geinin-ga gakuya-ni keeki-o sashiireshita fan-o hometa.
comedian-nom room-dat cake-acc brought fan-acc praised.
‘The comedian praised the fan who brought a cake to the dressing room.’
b. GP/ditransitive:
geinin-ga gakuya-ni keeki-o sashiireshita fan-o maneita.
comedian-nom room-dat cake-acc brought fan-acc invited.
‘The comedian invited the fan who brought a cake to the dressing room.’
c. non-GP/monotransitive:
26 Kusaba, S. and Yano, M.
shishoo-ga kinoo deshi-o hometa.
professor-nom yesterday pupil-acc praised.
‘The professor praised a pupil yesterday.’
d. non-GP/ditransitive:
sensei-ga shogakkoo-ni hogosha-o maneita.
teacher-nom E.S-dat parents-acc invited.
‘The teacher invited parents to the elementary school.’
A total of 60 sets of experimental sentences were distributed into two lists following the
Latin square method so that participants saw either a monotransitive or ditransitive verb in
each GP condition. The lists were counterbalanced among the participants.
Each sentence was presented in a phrase-by-phrase, non-cumulative manner. Participants
were asked to read a sentence phrase by phrase at a natural speed by pressing a space key.
To ensure that our participants were paying enough attention to the reading task, a yes-no
comprehension question was given following half of the trials of the GP conditions. Prior to the
experiment, eight practice trials were given to familiarise the participants with the self-paced
reading task. The experimental sentences were presented in a random order using LinguaTools.
Participants
Twenty native speakers of Japanese from Kyushu University participated in the
experiment (four males and 16 females, mean age = 21.4 years, range: 20–23). All participants
had normal or corrected-to-normal vision and no history of reading disability or neurological
or mental disorder. Written informed consent was obtained from all participants prior to the
experiment, and they were paid for their participation.
Data analysis
We analysed only the trials in which the comprehension question was answered correctly.
Reading time data exceeding 2.5 standard deviations from a participant’s mean at each region
were discarded. Statistical analysis was conducted using linear mixed-effects (LME) models
fitted with the lmer function of the lme4 package in R (Bates, Maechler, Bolker, & Walker,
2015). The models included the independent variables of interest (i.e., GP and VERB) as xed
factors. Each experimental condition was coded such that the GP and non-GP conditions were
assigned −0.5 and 0.5, respectively. Similarly, the monotransitive and ditransitive conditions
were assigned values of −0.5 and 0.5, respectively. The participants and items were treated
as random factors. The maximal model was built, as shown in (12) below, and then a final
model was selected using the backward stepwise method by comparing models using the
anova function of the lme4 package. P-values were calculated by submitting the nal model to
the lmer function of the lmerTest package (Kuznetsova, Brockhoff, & Christensen, 2015). An
interaction of GP by VERB was decomposed by conducting separate analyses at each GP type.
27
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
(12) Model = lmer (RT ~ GP * VERB + (1 + GP * VERB | Subj) + (1 + GP * VERB |
Set) + TrialOrder, data = data)
Prediction
Assuming that our participants prefer a simple sentence analysis upon reaching the rst
verb (i.e., ‘brought’) in the GP condition, they have to revise this analysis to build a structure in
(13a) or (13b) after encountering the head noun of the relative clause (i.e., ‘fan-acc’).
(13) a. [[comedian-nom] [gap1 room-dat cake-acc brought fan-acc 1] ... ]
b. [[comedian-nom] [room-dat] [gap1 cake-acc brought fan-acc1] ... ]
The MMPP requires the parser to build a maximal projection headed by an error signal
(i.e., ‘fan-acc’) with the fewest nodes upon encountering the error signal. Accordingly, this
principle predicts that the parser would select (13b), in which the error single (‘fan-acc’) does
not include a dative phrase (i.e., ‘room-dat’) in the relative clause.
6
The structure in (13b) is
compatible with a ditransitive verb, such as ‘invited’, but not with a monotransitive verb, such
as ‘praised’. Therefore, the MMPP predicts a longer reading time at the matrix verb for the
GP/monotransitive condition compared to the GP/ditransitive condition, reecting a second
structural reanalysis.
On the other hand, the MRP proposes that the parser holds an existing structure as much
as possible to incorporate an error signal into it. Accordingly, (13a) should be constructed
based on this principle. A monotransitive verb such as ‘praised’ can complete this existing
structure in a legitimate way. In contrast, a ditransitive verb requires the parser to revise an
disambiguated structure again by positing an empty category at the dative NP position or by
reanalysing ‘room-dat’ as a GOAL of the matrix verb ‘invited’. Consequently, the ditransitive
verb should incur a longer reading time than the monotransitive verb in the GP condition if
the parser adopts the MRP.
The difference between the monotransitive and ditransitive verb in the non-GP conditions
was not predicted by the MMPP or MRP. In other words, only the interaction between GP and
VERB was of interest to test these two hypotheses.
Results
Figure 1 shows mean reading times at each region, and Figure 2 shows mean reading times
at the matrix verb (i.e., the sixth region in the GP conditions and the fourth region in the non-
GP conditions). The linear mixed-effects model of reading times at the verb region showed a
marginally signicant effect of GP due to a longer time for the GP condition than the non-GP
condition (Table 1). The interaction did not reach signicance, although the reading time of the
GP/ditransitive condition was numerically longer than that of the GP/monotransitive condition.
6. Following Oishi (2007), we do not posit a gap for the GOAL phrase.
28 Kusaba, S. and Yano, M.
400
500
600
700
800
900
R1 R2 R3 R4 R5 R6
Reading mes (ms)
GP/monotransitive
GP/ditransitive
non-GP/monotransitive
non-GP/ditransitive
Figure 1. Mean reading times at each region. Error bars indicate standard errors.
687.9
719.9
640.9 653.6
500
600
700
800
GP/monotransitive GP/ditransitive non-GP/monotransitive non-GP/ditransitive
Reading mes (ms)
Figure 2. Mean reading times (ms) at the verb region. Error bars indicate standard errors.
Table 1. Summary of the xed effects in the linear mixed-effects
(LME) model of reading times at the verb region.
t p
For the comprehension question task, the effect of VERB was marginally signicant in
the response times and signicant in accuracy (Figure 3 and Table 2).
7
This result showed that
7.
As noted above, the comprehension questions were given only for the GP conditions. Hence, the effect of
interest was only an effect of VERB in the comprehension question task. The length (the number of morae)
was matched between the GP/monotransitive and GP/ditransitive conditions (t (62) = −0.28, p > 0.10).
29
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
the participants took more time to answer a comprehension question and that their answers
were less accurate in the GP/ditransitive condition than in the GP/monotransitive condition.
90.6
75.3
70
80
90
100
GP/monotransitive GP/ditransitive
Accuracy (%)
2397.1
2852.7
2200
2400
2600
2800
3000
3200
GP/monotransitive GP/ditransitive
Response times (ms)
Figure 3. Mean response times (left) and accuracy (right) in the comprehension question
task. Error bars indicate standard errors.
Table 2. Summary of the xed effects in the linear mixed-effects (LME) model of response times (left)
and accuracy (right) in the comprehension question task.
t p z p
Discussion
In Experiment 1, we failed to find an interaction of GP by VERB at the verb region,
which was not predicted by the MMPP or the MRP. Since our critical region was at the end
of a sentence, the lack of interaction may be due to a delayed effect of processing cost, which
has often been observed in the self-paced reading task. In other words, the processing cost
may increase after participants nished reading a sentence, and thus, we could not measure
it successfully. In Experiment 2, we tested our prediction using a different task to avoid the
delayed effect.
Experiment 2
Experiment 2 employed a speeded yes-no acceptability judgement task to detect
processing costs at the end of a sentence. This task required participants to judge the
acceptability of a sentence as quickly as possible. We assume that when participants need
to reanalyse a syntactic structure upon encountering a matrix verb, they should experience
a processing difficulty, and their responses should be delayed accordingly. Thus, by using
response times as an index of reanalysis cost, we can assess the syntactic structure into which
participants disambiguate a temporally ambiguous structure.
30 Kusaba, S. and Yano, M.
Stimuli
The same stimuli used in Experiment 1 were employed for the target sentences (i.e., YES
responses). A total of 60 unacceptable sentences for NO responses were added into each list.
Unacceptable sentences included case-assignment violation or semantic anomaly between a
verb and its argument. Half of the ller sentences consisted of six phrases including a relative
clause, and the other half consisted of four phrases without a relative clause. The lists were
counterbalanced among the participants.
After a xation was presented for 700 ms, each phrase was presented at the centre of the
screen for 700 ms with an inter-stimulus interval of 100 ms. The participants were instructed
to judge the acceptability of a sentence and press YES (acceptable) or NO (unacceptable)
buttons as soon as possible after the matrix verb and a period appeared.
8
After every ve trials
in the GP condition, a comprehension question was given to check whether the participants
could understand the sentences correctly. The experimental sentences were presented in a
random order using Presentation 16.5 (Neurobehavioral Systems).
Participants
Sixteen native speakers of Japanese were recruited from Kyushu University (11 males and
ve females, mean age = 21.5 years, range: 20–24). All participants had normal or corrected-
to-normal vision and no history of reading disability or neurological or mental disorder.
Written informed consent was obtained from all participants prior to the experiment, and they
were paid for their participation.
Prediction
According to the MMPP, a temporarily ambiguous structure should be disambiguated
into the ditransitive structure with the fewest nodes inside of the relative clause headed by the
error signal. Hence, this principle predicts a longer response time for the GP/monotransitive
condition than for the GP/ditransitive condition. On the other hand, the MRP offers the
opposite prediction. If the parser follows this strategy, it should not reanalyse a dative NP
(‘room-dat’) and an accusative NP (‘cake-acc’) belonging to a different clause at the head
noun since this principle requires the parser to maintain as much of the existing structure as
possible when required to revise a structure (see (13)).
Results
The statistical analyses were conducted in the same way as in Experiment 1. Figures
4 and 5 show the mean response times and accuracy in the acceptability judgement task,
respectively. The result showed a signicant interaction of GP by VERB on response times
(Table 3). The planned comparison revealed a signicant effect of VERB in the GP condition
(
β
= 140.7, SE = 52.79, t = 2.66, p = 0.01) but not in the non-GP condition (
β
= −3.30,
8. In every type of sentence used in Experiment 2, a yes-no response was possible only after reading a
matrix verb. The response time refers to duration (ms) from the onset of the matrix verb to the response.
31
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
SE = 37.76, t = −0.09, p = 0.93). This result indicates that participants experienced greater
processing difculty in the GP/ditransitive condition than in the GP/monotransitive condition.
The interaction was also signicant for accuracy of the acceptability judgement (
β
= 1.73,
SE = 0.54,
z
= 3.18, p < 0.01). Again, the planned comparison found a signicant effect of
VERB only in the GP condition (GP:
β
= −1.24, SE = 0.30,
z
= −4.11, p < 0.01; non-GP:
β
= 0.51, SE = 0.46,
z
= 1.1, p = 0.26).
1003.5
1141.9
1005.8 999.5
900
1000
1100
1200
1300
GP/monotransitive GP/ditransitive non-GP/monotransitive non-GP/ditransitive
Response times (ms)
Figure 4. Mean response times in the acceptability judgement task. Error bars
indicate standard errors.
96.3
89.1
3.893.79
80
90
100
GP/monotransitive GP/ditransitive non-GP/monotransitive non-GP/ditransitive
Accuracy (%)
Figure 5. Mean accuracy in the acceptability judgement task. Error bars
indicate standard errors.
Table 3. Summary of the xed effects in the linear mixed-effects (LME) model
of response times in the acceptability judgement task.
t p
32 Kusaba, S. and Yano, M.
We also compared response times and accuracy in the comprehension question task
between the GP/monotransitive and GP/ditransitive conditions (Figure 6). The effect of VERB
did not reach signicance in the response time data, whereas the response was signicantly
more accurate in the GP/monotransitive condition than in the GP/ditransitive condition (Table
4). This result indicates that our participants computed their nal interpretation of the GP/
ditransitive sentence less accurately at the end of a sentence.
1565.3
1642.2
1400
1500
1600
1700
1800
GP/monotransitive GP/ditransitive
Response times (ms)
96.5
88.7
80
90
100
GP/monotransitive GP/ditransitive
Accuracy (%)
Figure 6. Mean response times (left) and accuracy (right) in the comprehension question task.
Error bars indicate standard errors.
Table 4. Summary of the xed effects in the linear mixed-effects (LME) model of response times (left)
and accuracy (right) in the comprehension question task.
t p
Discussion
The result of the speeded acceptability judgement task revealed a longer response time
for the GP/ditransitive condition than for the GP/monotransitive condition. This increased
response time for the GP/ditransitive verb should not reflect lexical differences of verbs,
such as lexical frequency and familiarity that are known to modulate lexical access costs,
because there was no difference between the non-GP/ditransitive and non-GP/monotransitive
conditions. The accuracy in the GP/ditransitive condition was lower than that in the other
three conditions. These results indicate that the participants faced processing difculty when
reading a ditransitive verb in the GP condition, suggesting a second reanalysis cost in this
condition. In other words, this result indicates that when reading an unexpected head noun,
the parser disambiguated a temporally ambiguous structure into a monotransitive structure,
such as in (14a) below, rather than a ditransitive structure, as in (14b). This nding is in favour
of the prediction based on the MRP. This principle predicts that the participants did not
need to revise a structure again when reading monotransitive verbs, such as ‘praised’, since
33
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
(14a) is compatible with it. However, a ditransitive verb such as ‘invited’ forced participants
to reanalyse ‘room-dat’ as a GOAL argument or posit a GOAL that was not mentioned
in the sentence. Thus, this principle successfully predicts a higher processing cost for the
GP/ditransitive condition, which is consistent with our results. In contrast, if the parser
disambiguated an ambiguous structure into the ditransitive one in (14b), as expected by
the MMPP, the parser would have needed to reanalyse ‘room-dat’ as a constituent inside the
relative clause when reading a monotransitive verb, as the monotransitive verb cannot take
three arguments (‘comedian-nom’, ‘room-dat’, and ‘fan-acc’). Accordingly, the monotransitive
verb should have incurred a higher processing load than the ditransitive condition. Because
this expectation is contrary to the processing asymmetry of the monotransitive and
ditransitive verbs in Experiment 2, this principle cannot offer a plausible explanation for our
results.
(14) a.
[[comedian-nom] [gap1 room-dat cake-acc brought fan-acc1] ... ]
b. [[comedian-nom] [room-dat] [gap1 cake-acc brought fan-acc 1] ... ]
General Discussion
The purpose of this study was to explore an ambiguity resolution mechanism in the
process of reanalysis by testing the MRP and the MMPP. The result of the self-paced reading
experiment neither supported the MRP nor the MMPP due to the lack of a signicant effect
at the sentence-nal matrix verb region. This may be due to a delayed effect of the processing
cost induced by the matrix verb. However, the response times for the comprehension question
gave us a hint as to the disambiguation preference for the monotransitive structure over the
ditransitive one. In Experiment 2, we performed a speeded yes-no acceptability judgement
experiment to avoid a delayed effect and measured the processing cost pertaining to a second
reanalysis. The results of Experiment 2 showed a processing advantage for a monotransitive
structure over a ditransitive one only when a sentence fragment could be analysed as either
structure in the ambiguity resolution (i.e., GP sentences). Our evidence argues for the MRP,
but not for the MMPP.
9
However, it is fair to say that the MMPP can account for the result
obtained by Oishi (2007), but the MRP cannot. To summarise, the MRP cannot account for
the result of Oishi’s (2007), whereas the MMPP cannot account for the result of Experiment 2
of the present study. We propose a new account below that correctly predicts a disambiguation
preference in the present and previous experiments.
Here, we summarise relevant empirical data obtained from Oishi (2007) and the present
study before going to the details of the new account. (15a) shows an initial analysis of a
sentence before facing the first GP effect. (15b) and (15c) show available disambiguated
structures in which (15b) is compatible with a ditransitive verb and (15c) is compatible with
a monotransitive verb. According to Oishi (2007), the monotransitive verb incurred a larger
processing cost than the ditransitive verb, suggesting that the parser selected the structure in
34 Kusaba, S. and Yano, M.
(15b). This structure includes fewer nodes within the relative clause headed by an error signal
of bribe-acc’ in comparison to the one in (15c). Therefore, his result supported the MMPP.
(15) Oishi (2007):
a. daijin-ga honbu-ni atsumatta
minister-nom headquarters-dat gathered
‘The minister(s) gathered at the headquarters’
minister-NOM
(entrusted)
headquarters-DAT
gap
i
gathered
bribei-ACC
b.
(stole)
minister-NOM
bribei-ACC
gathered
gapi
headquarters-DAT
c.
The example (16a) shows an initial analysis of the GP sentence used in Experiments 1 and
2. The examples (16b) and (16c) show ditransitive and monotransitive structures, respectively.
The MMPP predicts that (16b) should be selected out of these two options for the same reason
mentioned above. However, our ndings suggest that the parser built the structure in (16c),
evidenced by the longer response times and less accurate responses for the ditransitive verb
than for the monotransitive verb. To summarise, the Japanese parser builds a ditransitive
structure when encountering an error signal preceded by an intransitive verb, whereas it
constructs a monotransitive structure when encountering an error signal preceded by a
transitive verb.
9. One might think that if a dative GOAL phrase is analysed as an argument rather than an adjunct of
brought’, the MMPP successfully predicts a GP effect in the present study, since this principle states
that the parser needs to ll an argument position with an overt constituent if possible. Thus, the dative
phrase ‘room-dat’ should be located inside the relative clause. However, if this view is correct, it is not
obvious why the dative GOAL phrase ‘headquarters-dat’ should be analysed as a constituent outside
of the relative clause with ‘gathered vi’, which is also supposed to take a GOAL argument in Oishi’s
(2007) experiment. Since the argument/adjunct distinction is not clear-cut, this issue awaits further
investigation.
35
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
(16) Experiments 1 and 2
a. geinin-ga gakuya-ni keeki-o sashiireshita
comedian-nom room-dat cake-acc brought
‘The comedian brought the cake to the dressing room’
comedian-NOM
(invited)
room-DAT
fani-ACC
brought
gapi
cake-ACC
b.
comedian-NOM
(praised)
fani-ACC
room-DAT
gapi
brought cake-ACC
c.
As shown in (17) below, the common feature of the disambiguated structure in Oishi’s
(2007) study and the present experiment is that the initially analysed clause-mate relationship
between a THEME phrase and a GOAL phrase holds even after incorporating an error
signal into an initial structure. In the case of Oishi’s (2007) experiment, ‘minister-nom’ and
headquarters-dat’ are analysed as clause-mate constituents belonging to the matrix clause. On
the other hand, the parser analyses ‘room-dat’ as a constituent inside of the relative clause to
maintain the clause-mate relationship with the following ‘cake-acc’. Therefore, when required
to revise a syntactic structure of a sentence, the Japanese parser favours positing NPs that
are initially analysed as the THEME and GOAL as clause-mate constituents governed by the
same verb. This hypothesis can correctly predict a GP effect in both Oishi’s (2007) experiment
and the present research.
36 Kusaba, S. and Yano, M.
(17)
Conclusion
We investigated structural preference pertaining to ambiguity resolution in the process of
reanalysis during Japanese sentence comprehension through self-paced reading and speeded
yes-no acceptability judgement experiments. Our findings favour the Minimal Revision
Principle, which assumes that the parser should avoid unnecessary structural analysis to
the extent possible. In contrast, our evidence is not consistent with the Minimum Maximal-
Projection Principle, which posits that the parser builds the smallest maximal projection
headed by an error signal. We proposed a new account following the spirit of the MRP to
provide a consistent explanation for the present and previous experiments. That is, the
Japanese parser prefers to maintain an initial analysis of THEME and GOAL phrases as being
constituents of the same clause upon encountering an error signal (i.e., head noun).
Acknowledgement
This study was supported by a Grant-in-Aid for JSPS Research Fellows from the Japan
Society for the Promotion of Science (#13J04854, PI: Masataka Yano). We are grateful to
Tsutomu Sakamoto and Masatoshi Koizumi for their insightful comments. We would also like
to thank Yu Bise, Yuki Tateyama, and Yoan Kim for their assistance with the data collection.
This study is based on a B.A. thesis submitted by the first author to the Department of
Linguistics, Kyushu University.
References
Amano, S., & Kondo, T. (Eds.). (1999). Nihongo-no goitokusei [Lexical Properties of Japanese]. NTT Communication
Science Laboratories, Sanseido, Tokyo: Japan.
minister-NOM
(entrusted)
headquarters-DAT
gapi gathered
bribei-ACC
comedian-NOM
(praised)
fani-ACC
room-DAT
gapi
brought cake-ACC
37
Structural ambiguity resolution in the process of reanalysis: Evidence from Japanese sentence comprehension
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). lme4: Linear mixed-effects models using Eigen and S4. R
package version 1.1-8.
Bever, T. G. (1970). The cognitive basis for linguistic structures. In J. R. Hayes (Ed.), Cognition and the Development
of Language, 279–362, New York: Wiley & Sons.
Ferreira, F., & Clifton, C., Jr. (1986). The independence of syntactic processing. Journal of Memory and Language,
25(3), 348–368.
Frazier, L. (1979). On comprehending sentences: syntactic parsing strategies. Doctoral Dissertation, University of
Massachusetts.
Frazier, L., & Clifton, C., Jr. (1998). Sentence Reanalysis, and Visibility. In J. D. Fodor, & F. Ferreira (Eds.),
Reanalysis in sentence processing, Dordrecht, The Netherlands: Kluwer Academic Publishers, 143–176.
Frazier, L., & Fodor, J. D. (1978). A sausage machine: A new two-stage parsing model. Cognition, 6(4), 291–325.
Hirose, Y., & Inoue, A. (1998). Ambiguity of reanalysis in parsing complex sentences in Japanese. Syntax and
semantics, 31, 71-93.
Inoue, M. (2006). Nihongobun-no rikai-niokeru aimaisei-no kaisho-to horyuu [Ambiguity resolution and retention in
Japanese sentence comprehension]. Cognitive Studies, 13(3), 353–368.
Kuznetsova, A., Brockhoff, P. B., & Christensen, R. H. B. (2015). lmerTest: Tests in linear mixed effects models. R
package version 2.0-29.
Mazuka, R., & Itoh, K. (1995). Can Japanese speakers be led down the garden-path? In R. Mazuka, & N. Nagai (Eds.)
Hillsdale, NJ: Lawrence Erlbaum, Japanese sentence processing, 295–329.
Oishi, H. (2007). Saibunseki-niokeru ichijitekiaimaisei-no kaisho-katei-nituite: Tajuujoohookan-no kyoogoo-no
kanoosei [On temporal ambiguity resolution process in reanalysis: Possibility of competition between
multiple (sources of) information]. Doctoral Dissertation, Kyushu University.
Oishi, H., Yasunaga, D., & Sakamoto, T. (2007). Revision process in Japanese sentence processing: Evidence from
event-related brain potentials. In T. Sakamoto (Ed.), Communicating Skills of Intention, 367–381, Tokyo:
Hitsuji Syobo.
Sturt, P., & Crocker, M. W. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and
reanalysis. Language and Cognitive Processes, 11(5), 449–494.
(Received September 20, 2017)
(Accepted December 14, 2017)
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
One of the frequent questions by users of the mixed model function lmer of the lme4 package has been: How can I get p values for the F and t tests for objects returned by lmer? The lmerTest package extends the 'lmerMod' class of the lme4 package, by overloading the anova and summary functions by providing p values for tests for fixed effects. We have implemented the Satterthwaite's method for approximating degrees of freedom for the t and F tests. We have also implemented the construction of Type I - III ANOVA tables. Furthermore, one may also obtain the summary as well as the anova table using the Kenward-Roger approximation for denominator degrees of freedom (based on the KRmodcomp function from the pbkrtest package). Some other convenient mixed model analysis tools such as a step method, that performs backward elimination of nonsignificant effects - both random and fixed, calculation of population means and multiple comparison tests together with plot facilities are provided by the package as well.
Article
Full-text available
Thesis - University of Connecticut. Photocopy.
Article
Full-text available
This paper describes a new addition to the family of parsing models based on the principles of Description-Theory (Marcus, Hindle & Fleck, 1983). We demonstrate how the definition of two simple parsing operations, simple attachment and tree lowering, which are related to the grammatical composition operations of substitution and adjunction in the Tree Adjoining Grammar formalism, yields a parser which is more constrained than previous Description-theory based models (e.g. Gorrell 1995). Since the tree-lowering operation allows the parser to reanalyse in the case of "unconscious" garden paths, it can be used to investigate the consequences of adopting various search strategies for reanalysis, predicting preferences in cases where more than one possibility for reanalysis exists. Considering data from English and Japanese, we show that reanalysis preferences may differ between head-final and head-initial languages, and suggest some reasons why this might be so. 1 Introduction A psychologi...
Chapter
Serial theories of sentence processsing specify that the processor pursues just a single syntactic analysis of a sentence until that analysis becomes implausible or untenable, at which point revision of the first analysis occurs. Many theories of revision have been proposed in recent years. Here it is argued that revision cost cannot be calculated in purely structural terms: reanalysis complexity is a function of sentence token, not sentence type. It is also argued that a theory of revisions must include a Minimal Revisions principle. Finally, in revision and in first analysis, more recently postulated nodes provide preferred or more “visible” attachment sites than more distant nodes. Several types of evidence are used to support a Visibility principle which, it is suggested, may replace principles such as Late Closure (Frazier, 1978) or Recency (Gibson, 1991).
Article
It is proposed that the human sentence parsing device assigns phrase structure to word strings in two steps. The first stage parser assigns lexical and phrasal nodes to substrings of roughly six words. The second stage parser then adds higher nodes to link these phrasal packages together into a complete phrase marker. This model of the parser is compared with ATN models, and with the two-stage models of Kimball (1973) and Fodor, Bever and Garrett (1974). Our assumption that the units which are shunted from the first stage to the second stage are defined by their length, rather than by their syntactic type, explains the effects of constituent length on perceptual complexity in center embedded sentences and in sentences of the kind that fall under Kimball's principle of Right Association. The particular division of labor between the two parsing units allows us to explain, without appeal to any ad hoc parsing strategies, why the parser makes certain 'shortsighted' errors even though, in general, it is able to make intelligent use of all the information that is available to it.
Article
Three experiments addressed the question whether semantic content or pragmatic context can direct the initial syntactic analysis assigned to sentences. Each experiment determined whether syntactic processing biases that have been observed in sentences presented in isolation can be overcome. In two experiments that measured eye movements, we found that the syntactic processing biases remained even when they resulted in thematically based anomaly or when they conflicted with discourse biases. In a third experiment, we used a self-paced reading task to replicate some of the results obtained using eye movement measures. We argue that the data support the existence of a syntactic processing module.